Mission Control

Mission Control — Fleet Dashboard

Open-source agent orchestration dashboard (Next.js 16). Runs on Mac Mini, accessible fleet-wide via Tailscale.

Quickstart

cd /Users/appie/mission-control
pnpm run start --port 3480

If pnpm run start fails on verify:node or pnpm install checks, bypass it:

npx next start --hostname 0.0.0.0 --port 3480

Access

Dashboard: http://100.101.29.56:3480 (Tailscale IP of Mac Mini) Default creds: see references/credentials.md

Pitfalls

0. AUTH_USER/AUTH_PASS from .env doesn't propagate to database on first run

Root cause: Setting AUTH_USER and AUTH_PASS in .env seeds the admin credentials into MC's config but does NOT create a user record in the database. The /api/login endpoint checks the DB, not the env vars, so returns 401 even with correct creds.

Fix: Visit http://localhost:3480/setup in a browser on first run to complete the admin account creation wizard. The env vars only take effect if the setup has already seeded from them — if the setup page was never visited, no DB user exists.

Workaround (headless): Post to /api/setup endpoint with the same credentials to create the DB record.

1. External access returns 403 Forbidden

Root cause: MC_ALLOWED_HOSTS in .env defaults to localhost,127.0.0.1,::1. In production (NODE_ENV=production), the proxy (src/proxy.ts:163) blocks any host not in the allowlist.

Fix: Add Tailscale and local network ranges to .env:

MC_ALLOWED_HOSTS=localhost,127.0.0.1,::1,100.*,192.168.*

Patterns supported: exact match, *.example.com (subdomain), 100.* (prefix wildcard).

Then restart the server — .env changes are read at startup, not hot-reloaded.

2. pnpm verify:node fails

pnpm run start runs verify:node first which may fail on pnpm install checks. Workaround: skip pnpm entirely with npx next start --hostname 0.0.0.0 --port 3480.

3. Server process not found after restart

Check what's listening:

lsof -i :3480

Find project dir of running process:

lsof -p <PID> | grep cwd
# or:
ps aux | grep <PID>
# gives path like: /Users/appie/mission-control

4. DeepSeek models fail on vision/image tools

DeepSeek V4 Pro (and likely other DeepSeek models) return 404 - No endpoints found that support image input. For vision tasks, use anthropic/claude-opus-4-6 or another vision-capable model.

Connecting session tasks/context to Mission Control

Mission Control's agent task surface is the bundled Hermes Kanban dashboard plugin at /kanban, backed by hermes_cli/kanban_db.py. When Seyed asks to put tasks/context into Mission Control, mirror the active session objective into the Hermes Kanban board rather than only keeping an in-chat todo.

Recommended flow:

1. Inspect boards with ./hermes kanban boards list from the Hermes Agent repo, or hermes kanban boards list if the CLI is on PATH.
2. Create a task with a concise title and rich --body containing: user request, current repo/workdir, active goal/session identifier if available, relevant constraints, and a short checklist of mirrored tasks.
3. Use numeric priority values, not labels. Example: --priority 80, not --priority high.
4. Prefer --workspace dir:<absolute-path> when the context belongs to an existing checkout.
5. Verify with ./hermes kanban list and ./hermes kanban show <task-id> so Mission Control has the expected body.

Pitfall: python -m hermes_cli.kanban ... may have no CLI entrypoint in repo checkouts. Use the Hermes CLI command path instead.

Reference: references/session-context-kanban.md has a concrete command template for mirroring session context into Mission Control.

For fleet-wide context ingestion across Appie profiles, use references/fleet-context-ingestion.md: it lists authoritative evidence files, card body shape, idempotency keys, and verification steps. Prefer a dedicated board such as appie-fleet-context over dumping cross-profile cards into the default board.

Diagnostics

Check Diddy/Appie Hermes logs

# Recent errors
tail -50 ~/.hermes/logs/errors.log
# Gateway errors
tail -20 ~/.hermes/logs/gateway.error.log
# Agent activity
tail -30 ~/.hermes/logs/agent.log

Common patterns to watch for:

• unknown provider 'openai' — vision provider not configured; falls back to auto
• No endpoints found that support image input — current model can't handle images
• Telegram network error — connectivity issue (usually transient)
• Discord RESUMED session — normal if occasional; if every 2-3h, gateway may be unstable

References

• references/access-patterns.md — full MC_ALLOWED_HOSTS proxy logic and access control
• references/log-patterns.md — common Hermes log warnings and their meanings
• references/fleet-topology.md — current Appie fleet layout (local vs remote agents)